FDA Uses AI to Inspect QMS, MHRA Demands Continuous AI Oversight

On 6 May 2026, FDA upgraded its internal AI tool Elsa to version 4.0 and connected it to HALO, a data platform consolidating over 40 submission and application sources. The result is not faster review. It is total visibility. Sampling is no longer the limit during inspections. FDA can now interrogate your entire quality management system, trace every CAPA back to its root cause analysis, and surface patterns you may not have seen yourself. Eleven months after Elsa's June 2025 launch, the agency has moved from pilot to production. Meanwhile, MHRA has adopted 44 recommendations for AI-enabled medical devices, including continuous oversight throughout the device's working life. For device manufacturers, these are not parallel developments. They are the left and right edges of a regulatory envelope closing around quality systems and post-market surveillance architecture.
What FDA's Elsa 4.0 Actually Does
Elsa is a generative AI tool built for FDA staff, not industry. Its integration with HALO means inspectors can query across premarket submissions, post-market reports, establishment registrations, adverse event data, and recall histories in a single interface. The practical consequence is that inspectors arrive at your facility already knowing where the gaps are. They can correlate a 510(k) claim with MDR trends, link a design change to a CAPA closure, and flag inconsistencies between your submission narrative and your actual device history record. This is not speculation. FDA has been using Elsa in live inspections since mid-2025, and the upgrade to 4.0 expands both the dataset and the sophistication of the queries.
The shift is from sample-based inspection to system-level interrogation. Historically, inspectors reviewed a subset of records during a site visit. Now they can examine the entire population, identify outliers, and drill into specific device history files before they walk through your door. The implication for quality teams is straightforward: your QMS must withstand algorithmic scrutiny, not just human sampling. If your CAPA system closes tickets without linking corrective actions to verification evidence, Elsa will find it. If your design controls show version jumps without corresponding change orders, the pattern will surface. The bar is no longer regulatory minimalism. It is structural integrity across the entire quality system.
This connects directly to the FDA's broader move toward AI-enabled compliance architecture, where speed and oversight converge. For SaMD teams especially, the expectation is that your quality system can explain itself to an algorithm, not just an auditor.
MHRA's Continuous Oversight Mandate for AI Devices
MHRA has adopted 44 recommendations from its AI and medical devices working group, with continuous lifecycle monitoring as the centrepiece. The principle is that AI-enabled devices change in ways that traditional fixed-function devices do not. Model drift, retraining, performance decay, and edge-case failures are not hypothetical risks but operational realities. MHRA's position is that regulatory oversight cannot end at market approval. It must extend across the device's working life, with manufacturers required to monitor performance, report anomalies, and update risk assessments as the device evolves.
The recommendations do not yet specify the technical architecture for continuous monitoring, but the regulatory expectation is clear. Manufacturers must have systems in place to detect when device performance deviates from its validated state, to trace those deviations to root causes, and to intervene before harm occurs. This is not post-market surveillance as an afterthought. It is post-market surveillance as a design input, baked into the device architecture from the start.
| Dimension | FDA Elsa 4.0 | MHRA Continuous Oversight |
|---|---|---|
| Scope | Premarket and post-market data integration for inspection | Lifecycle monitoring mandate for AI-enabled devices |
| Primary mechanism | Algorithmic interrogation of quality system records | Manufacturer-operated performance tracking and reporting |
| Trigger | Routine inspections, targeted investigations | Continuous operation, anomaly detection |
| Data source | FDA internal databases (HALO platform) | Manufacturer post-market surveillance systems |
| Regulatory burden | Indirect: QMS must withstand AI scrutiny | Direct: Manufacturers must build and maintain monitoring infrastructure |
The Convergence: Quality Systems That Explain Themselves
The strategic implication is that quality management systems must now serve two masters. They must satisfy human auditors during site visits, and they must withstand algorithmic interrogation during inspections or continuous monitoring. The overlap is not complete. Human auditors care about narrative coherence, procedural adherence, and evidence of competence. Algorithms care about structural consistency, traceability, and the absence of unexplained gaps. The intersection is quality systems that are both procedurally sound and computationally legible.
For device manufacturers, this means rethinking how quality data is captured, linked, and stored. If your CAPA system lives in one database, your design history file in another, and your post-market surveillance data in a third, you have a problem. Elsa does not care about your organisational silos. It will correlate across them. MHRA's continuous oversight requirements will demand real-time performance data linked to risk assessments, with audit trails that prove you acted when performance decayed. The common denominator is structured, traceable, queryable quality data. Not PDFs. Not Word documents. Not spreadsheets with manual version control.
This is where quality management and regulatory expectations are converging. The architecture decisions you make today determine whether your QMS can scale with regulatory scrutiny or collapse under it.
What This Means for Your Team
If you are managing a quality system, the question is not whether FDA will use Elsa during your next inspection. It is whether your QMS can survive that interrogation. The practical test is simple: can you trace every post-market complaint back to a design input, through a risk assessment, into a verification protocol, and out to a closed CAPA with objective evidence of effectiveness? Can you do it algorithmically, not just manually? If the answer is no, you have a structural gap, not a documentation problem.
For AI-enabled device manufacturers, MHRA's continuous oversight mandate adds a second layer. You must now demonstrate that your device's performance is monitored in real time, that deviations are detected and assessed, and that corrective actions are triggered before failures reach patients. This is not a post-market surveillance bolt-on. It is a design requirement. If your device architecture does not include telemetry, logging, and anomaly detection as first-class features, you are not building to the regulatory expectation. You are building technical debt.
The operational consequence is that quality and regulatory teams must work with engineering from the start. Post-market surveillance architecture must be a design input, not a compliance afterthought. Risk management must extend beyond initial approval to continuous operation. And your QMS must be structured for algorithmic interrogation, not just human audit. The teams that recognise this early will have a compliance advantage. The teams that do not will spend the next two years retrofitting quality systems that were never designed to withstand this level of scrutiny.
Regulatory Scrutiny Pathway for AI-Enabled Devices
Key Takeaways
- FDA's Elsa 4.0 can interrogate your entire quality system, not just a sample. Inspectors arrive knowing where the gaps are, and your QMS must withstand algorithmic scrutiny.
- MHRA mandates continuous oversight for AI-enabled devices, with manufacturers required to monitor performance, detect anomalies, and update risk assessments throughout the device's working life.
- Quality systems must now serve two masters: human auditors and algorithmic interrogation. That demands structured, traceable, queryable data, not PDFs and spreadsheets.
- Post-market surveillance for AI devices must be a design input, not a compliance afterthought. If your architecture does not include telemetry and anomaly detection, you are building technical debt.
The regulatory envelope is closing. FDA can see your entire quality system, and MHRA expects you to watch your device in real time. For manufacturers, the question is no longer whether to invest in compliance architecture. It is whether you can afford to delay. The teams that build for algorithmic scrutiny now will have a structural advantage when the next inspection or surveillance audit lands. The teams that wait will be retrofitting under pressure, with regulators already inside the system.